Sam Altman’s OpenAI Defeats Elon Musk’s Grok in A.I. Chess Championship

On Thursday, OpenAI's o3 model—decommissioned last weekend with the release of GPT-5—defeated Elon Musk's xAI Grok 4 in four consecutive matches at Google's Kaggle chess competition.

One might assume this was an intricate high-tech showdown pushing reasoning capabilities to extremes, but let's start with an analogy from world chess champion Magnus Carlsen who described the AI matchup as "talented children who don't know how chess pieces move."

The three-day event from August 5-7 challenged general-purpose chatbots—those helping you draft emails and claiming near-human intelligence—to play chess without specialized training. No chess engines, no move lookups, just random chess knowledge absorbed from internet data.

Results resembled typical language model behavior in board games. During final commentary, Carlsen estimated both AIs operated at casual player levels (approx. 800 ELO), while his 2839 ELO as historical chess champion highlighted their deficiencies. Their gameplay seemed akin to learning from corrupted PDFs.

"They oscillate between brilliant moves and incomprehensible sequences," Carlsen remarked in a post-match livestream. At one point, he joked Grok might be playing "king of the hill" instead of chess after moving its king into immediate danger.

Even for non-chess audiences, these matches offered masterclasses in "how not to play chess." In the first game, Grok sacrificed key pieces gratuitously and worsened its position through excessive exchanges.

The second match grew stranger when Grok attempted a "poisoned pawn" strategy—an aggressive yet legal tactic where you capture what seems like a free enemy pawn that isn't. It mistakenly targeted a protected pawn, losing its queen immediately in the process.

By the third match, Grok established seemingly solid positioning with strong board control and no obvious threats. However, mid-game it made fatal mistakes, consecutively losing critical pieces.

This performance was particularly striking since pre-o3 Grok had shown considerable potential, earning praise from chess grandmaster Hikaru Nakamura who called it "objectively the best so far."

The fourth and final match provided actual suspense when OpenAI's o3 made a major early error in what would typically be decisive. Nakamura noted o3 still had "several small tricks" despite its disadvantage.

He proved correct as o3 reclaimed its queen and gradually squeezed victory while Grok's endgame collapsed like wet cardboard.

"Grok made numerous errors while OpenAI didn't," Nakamura concluded in his livestream. This outcome directly contradicted earlier sentiments about AI capabilities.

Timing proved disastrous for Elon Musk, who recently downplayed xAI's chess abilities on X, claiming it was merely an "incidental" skill with minimal development focus. The competition results proved this understatement.

In an earlier unofficial tournament earlier this year, international master Levy Rozman hosted a similar event with less advanced models. The situation turned chaotic with illegal moves, piece summoning errors, and miscalculations. Dedicated chess AI Stockfish eventually defeated ChatGPT. In a previous semi-final, Altman's AI also beat Musk's AI 2-0.

This competition differed significantly. Each bot received four chances to make legal moves before automatic disqualification. Early rounds saw AI attempting piece teleportation, resurrecting captured pieces, and moving pawns sideways like invented chess variants.

Several AI teams were disqualified.

Google's Gemini secured third place by defeating another OpenAI model, partially redeeming organizers. The bronze medal match featured an absurd draw where both AIs simultaneously held winning positions yet couldn't conclude the game.

Carlsen noted AIs excelled at calculating captured material but lacked winning strategies—like being great at collecting ingredients but incapable of cooking.

These are the same AI models corporate executives claim near-human intelligence, threaten white-collar jobs, and will revolutionize work processes. Yet they struggle to play a 1500-year-old board game without rule violations.

For now, we can rest assured AI won't control humans—at least not anytime soon.